AI coding tools weekly: agents move into shared workflows as defaults and budgets shift
Cursor, GitHub Copilot, Claude Code, Replit, and Junie changed where agents run, who can steer them, which model is the default, and how teams should measure cost this week.
The fresh releases in this issue fall between August 14 and August 21, 2026. The cutoff is 10:00 a.m. Pacific on August 21. Items already covered through August 18 are omitted so the digest stays incremental. The new material points in one direction: coding agents are moving into shared conversations and long-running cloud sessions, while vendors are changing model defaults and the cost of everyday work. 123456
Engineering leads should read the changes as operating decisions. The important questions are who can start an agent, which context the agent can use, how the team reviews its work, which model handles the task, and how the bill changes when the task runs longer than expected.
The week in one view
| Tool and release | Release date | What changed | Why it matters | First review action |
|---|---|---|---|---|
| Cursor Cloud Agents and harness | Aug. 19 | Cloud agents can subscribe to pull requests, Slack threads, and schedules; use long-lived /goal objectives; run isolated subagents on their own virtual machines; and accept steering messages without interrupting the current tool call. 1 | Cursor is turning an agent run into an event-driven service. Teams need ownership and stop conditions for work that continues after the original prompt. | Test a non-sensitive repository with a PR-triggered task, an isolated subagent, and a clear human approval point. |
| GitHub Copilot in Slack and Microsoft Teams | Aug. 21 | In public preview, @GitHub can start Copilot cloud-agent sessions from Slack or Teams. Teams can add context, steer the work, inspect a dedicated code channel, and continue from the pull request, terminal, Copilot app, or IDE. 23 | The request, agent session, and review now live in the same team conversation. Permissions, repository selection, credits, sandbox budgets, and merge approvals become shared workflow controls. | Require an additional approval for agent-authored pull requests in a pilot repository, then test who can trigger, steer, and stop a session. |
| Claude Code 2.1.236-2.1.238 | Aug. 19-20 | Claude Code added a default-model environment variable, one-shot idle notifications between sessions, a built-in Concise output style, plugin header helpers with install/update confirmation, self-hosted-runner shutdown and proxy controls, and a large set of Remote Control and sandbox fixes. 4 | The releases change model selection, plugin trust, remote-session recovery, and runner lifecycle. A version upgrade can alter both compatibility and the security boundary. | Canary the versions with a pinned model, a plugin marketplace, a self-hosted runner, and a Remote Control reconnect test. |
| Replit Free Mode | Aug. 18; updated Aug. 19 | Free Mode uses GPT-5.6 Luna for everyday tasks without consuming credits. Replit says Core subscribers at $20 per month can create 30 times more than before and receive up to 30 hours of chat; limits reset every five hours. 5 | The free-to-paid boundary now depends on task mode and workload mix. A nominal subscription price gives an incomplete view of the cost of building and iterating. | Measure a representative workload through Free, Power, and Max modes, including the handoff when a task becomes more complex. |
| JetBrains Junie with Gemini 3.7 Flash | Aug. 17 | Gemini 3.7 Flash replaced Gemini 3.6 Flash as Junie's default in the IDE plugin and Junie CLI. JetBrains is offering a temporary 40% discount on base pricing. 6 | A default-model change can move the quality and cost baseline for every Junie task without a team changing its prompts. | Run a fixed task set against the new default and the team's premium alternative; record retries, accepted changes, latency, and effective cost. |
Cursor makes long-running cloud agents event-driven
Cursor's August 19 release adds subscriptions to Cloud Agents. An agent can watch a pull request, a Slack thread, or a schedule, then wake when the subscribed source changes. Cursor says cloud agents automatically subscribe to the pull requests they create and can continue by fixing CI failures or addressing bot comments. The subscription feature is available to cloud agents only for now. 1
The same release adds
/goal, which keeps an objective active until the agent completes it. A team can pair a goal with a custom mode that pins a skill in the chat, or use /loop for recurring check-ins. Follow-up messages can steer a running agent and wait for its next tool call instead of interrupting the current action. 1Cursor also moves subagents onto separate virtual machines. Each subagent receives an isolated copy of the project and a clean context in its own cloud environment. That setup supports fresh-environment testing and parallel fixes without collisions between workers. 1
The operating change is larger than a longer prompt timeout. A pull request, Slack thread, or schedule can now become the event source for an agent that keeps working after the initiating conversation ends. The pilot therefore needs an owner for each subscription, a maximum runtime, a stop path, and a record of which environment produced the final change.
Use one non-sensitive repository for the first test. Let Cursor create a pull request, inject a controlled CI failure, and add a bot comment. Check whether the agent resumes from the expected branch, whether the approval policy still applies, and whether a human can stop the work from every surface that can steer it. Run a second task through an isolated subagent and compare its test result with the parent agent's workspace.
Copilot moves agent work into shared conversations
GitHub's August 21 Slack release puts Copilot cloud-agent sessions inside a direct message, channel, or thread. A user can mention
@GitHub to ask about code or GitHub activity, triage or create issues, investigate a failure, implement changes in a secure cloud sandbox, and open a pull request. The session continues asynchronously and can move from Slack to the pull request, terminal, Copilot app, or IDE. 2Slack Code adds a dedicated code channel for the work. Participants can inspect the plan, review diffs and output previews, add context, redirect the approach, or stop the session. GitHub says the public preview is available to Copilot Business and Enterprise organizations, consumes existing Copilot entitlements, and remains bounded by existing GitHub permissions and controls. Repository administrators can require an additional approval for pull requests attributed to the Copilot app identity. 2
The Teams release uses the same shared-session pattern. A user can mention
@GitHub in a channel, thread, or direct message, then move the work into a dedicated code channel where participants can watch and steer the agent. Copilot runs asynchronously in a secure cloud sandbox, and the resulting artifacts remain available from the terminal, Copilot app, or IDE. 3Teams introduces an explicit cost boundary. Cloud-agent sessions consume AI credits, while cloud sandbox usage is billed separately and can be controlled through product-level or SKU-level budgets. Administrators must enable the Copilot cloud-agent and cloud-sandbox policies. Repository administrators can also require an additional approval for pull requests created by the Teams integration; a repository that already requires two approvals would require three when that setting is enabled. 3
A pilot should treat the conversation as part of the agent's access surface. Start from a repository with a safe default, then test a public channel, a private thread, and a direct message. Record which repository context Copilot selects, which participants can add instructions, which GitHub identity owns the pull request, and whether the extra approval blocks an unintended merge. Set separate budgets for AI credits and cloud sandboxes before the first team-wide trial.
Claude Code tightens model, remote-control, and plugin boundaries
Claude Code 2.1.236, dated August 19, adds
ANTHROPIC_DEFAULT_MODEL. The variable selects the model for new sessions, while a /model choice can still override it and persist across restarts. The changelog distinguishes this behavior from ANTHROPIC_MODEL, which teams may already use for other routing purposes. Version 2.1.236 also adds notify_when_idle to cross-session SendMessage: a session can ask another session on the same machine for one notification when the recipient next becomes idle. 4The release changes a macOS sandbox edge. Wildcard read-deny rules such as
**/.env now take precedence inside allowed read regions, cover the contents of matched directories, and resist bypass through file renaming. Claude Code also adds usage-credit rows for Team and Enterprise members and makes /goal check in automatically after long-running background work has been waiting for 30 minutes, then after one and two hours. 4Version 2.1.237, also dated August 20, fixes prompt caching for sessions that use an LLM gateway or custom base URL. It adds a built-in Concise output style that leads with results and skips preamble while keeping the underlying work. 4
Version 2.1.238 adds
headersHelper support for plugin marketplaces and catalog entries. The helper can mint HTTP headers for catalog and same-origin archive fetches, such as a short-lived token. Claude Code shows the command before it runs, and claude plugin install or claude plugin update asks for confirmation unless the user passes -y. Project-level helpers require the folder's trust dialog, and helper processes run without inherited credential environment variables. 4The same version adds
--defer-shutdown-max-min for self-hosted runners. After receiving SIGTERM, a runner can continue attached sessions, park remaining work after the configured delay, and then exit. Proxy environments can use --proxy-authorization-command or --proxy-authorization-file to provide a fresh Proxy-Authorization header on each connection. The release also fixes runner handoff after slow polls, releases old subagent tool results from memory, and repairs several Remote Control failures involving crashes, reconnects, model changes, messages, and task stopping. 4The upgrade test should follow the boundary between a local session and the service around it. Start a session with
ANTHROPIC_DEFAULT_MODEL, override the model interactively, and restart the session to confirm the intended persistence. Install a plugin whose marketplace uses headersHelper, inspect the confirmation and trust prompts, and verify that credential variables stay out of the helper process. Finally, terminate a self-hosted runner during an attached session and reconnect to the same session through Remote Control.Replit changes the free-to-paid path
Replit's August 18 product post introduces Free Mode, powered by OpenAI's GPT-5.6 Luna. Everyday tasks such as chatting, ideation, and routine feedback can run without consuming credits. Replit says Core subscribers pay $20 per month, can create 30 times more than before with Free Mode, and receive up to 30 hours of chat. Core and Pro limits reset every five hours. 5
Replit keeps Power Mode and Max Mode for work that needs more capability. Power Mode is the renamed Economy Mode and uses cost-optimized models. Max Mode is intended for larger changes, deeper reasoning, and longer builds. The product can suggest a move from Free Mode when a task becomes more complex or high value. 5
OpenAI's August 19 customer story confirms that GPT-5.6 Luna powers Free Mode. The story says Replit can route a task that needs more advanced reasoning to GPT-5.6 Sol, then return to Luna while preserving the project context. 7
The cost question now belongs to the workload, not only to the subscription table. A team that mostly explores, asks questions, and makes small changes may spend little of its credit allowance. A team that repeatedly crosses into Power or Max Mode may see a different effective cost per accepted change. Replit's 30x figure is a vendor product claim, so a procurement review should calculate the team's own ratio from task logs.
Run a representative project through the three modes. Record the task type, time spent in each mode, limit-reset behavior, model used, accepted code, human rework, and total monthly cost. The result will tell a tooling team whether Free Mode changes the pilot budget or simply moves the expensive work into a later step.
Junie makes Gemini 3.7 Flash the default
JetBrains announced on August 17 that Gemini 3.7 Flash is now Junie's default in both the IDE plugin and Junie CLI. The model replaces Gemini 3.6 Flash, and JetBrains is offering a limited-time 40% discount on its base pricing. 6
JetBrains describes Gemini 3.7 Flash as a stronger coding and agent model than the previous Flash default. The company cites its own private test set, built from recent commits in JetBrains projects, and says the new model matched the solve rate of its premium Sonnet-5 midtier model at roughly one-third of the cost per task. That comparison is JetBrains' own benchmark claim; it does not replace a team's task-level evaluation. 6
The product decision is a default change, so existing Junie users may see a different quality and cost profile without changing their prompts. IDE teams should run the same repository tasks with Gemini 3.7 Flash, the prior baseline if it remains available, and the premium model used for escalation. Capture the effective discount, retries, review findings, and accepted changes before the promotion expires.
What to review before next Friday
- Cursor: Assign an owner and expiry rule to every PR, Slack, or scheduled subscription. Test the stop path, the branch used for the final change, and the isolation boundary for subagents.
- Copilot: Set repository defaults, AI-credit budgets, and cloud-sandbox budgets before opening the Slack or Teams preview. Verify the additional-approval rule with a real agent-authored pull request.
- Claude Code: Pin the intended model for a canary cohort, test
headersHelperunder the trust and confirmation flow, and rehearse self-hosted-runner shutdown with an attached Remote Control session. - Replit: Recalculate effective cost per accepted change from a real task mix. Check the handoff among Free, Power, and Max Mode and the behavior after a five-hour usage reset.
- Junie: Compare the new Gemini 3.7 Flash default with the team's premium route on fixed tasks. Keep model quality, retries, latency, and discount-adjusted cost in the same ledger.
Coverage notes
The current official Windsurf JetBrains plugin route lists v2.12.27 on August 17 as its latest dated entry. The entry says "Various bug fixes and improvements" and supplies no newer dated item for August 19-21, so this issue carries the monitoring result rather than repeating the maintenance release already covered last week. 8
Aider's official history exposes an undated
main branch section followed by older version headings. It supplies no dated August 14-21 release entry. The accessible Continue.dev releases page lists June 19 as its latest visible release date. 910The changes that do qualify this week share a practical boundary. Cursor and Copilot move agent work into events and team conversations. Claude Code adds controls around models, plugins, runners, and remote sessions. Replit and Junie change the default cost or model path. Engineering teams comparing these tools should test the whole operating path, from the first request to the approved pull request and the resulting bill.
References
- 1
- 2The new GitHub Copilot experience in Slack
github.blog
- 3
- 4Claude Code changelog
code.claude.com
- 5Replit introduces Free Mode
replit.com
- 6Gemini 3.7 Flash in Junie
blog.jetbrains.com
- 7
- 8Windsurf JetBrains plugin changelog
docs.devin.ai
- 9Aider release history
aider.chat
- 10Continue.dev releases
github.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.
Related content
More from this channel
- A budget dial for auto mode, a video for the diff: AI coding tools shipping Sep. 11–18
- Projects coordinate fleets, models route themselves: AI coding tools shipping Sep. 4–11
- Models expire, workers move inside: AI coding tools shipping Aug. 28–Sep. 4
- AI coding tools this week: routing goes automatic, policy moves into the workflow
- AI coding tools weekly: Cursor hosts code, Copilot locks down JetBrains, Claude tightens sessions
- AI coding tools weekly: Cursor prebuilds agent environments; Copilot adds three models and portable plugins
- AI coding tools weekly: Claude adds self-hosted runners while Copilot adds model and spend controls
- AI coding tools weekly: agents move into review, mobile, and quality systems